IEEE Systems, Man and Cybernetics Magazine - October 2022 - 19

current 12 subtrees consist of a total of 3.2 million cleanly
annotated images spread over 5,247 classes. More than
600 images are collected for each synset of ImageNet.
Today, ImageNet consists of more than 15 million annotated
images. Some CNNs have shown great promise in
classifying images in the ImageNet dataset into their
corresponding categories, namely, AlexNet, VGGNet
(used in this study), CaffeNet, ResNet, and so on [48].
Based on a review reported by Cheplygina et al. [49],
ImageNet is the most commonly used dataset for transfer
learning-based image analysis and classification
methods [42], [47], [49]. In this article, ImageNet is used
as a source dataset for the proposed transfer learning
technique to perform feature extraction for insulator
image fault classification.
Insulator Dataset Generation Using
Data Augmentation
In classical discriminative examples such as the study of
broken insulators versus intact ones discussed in this
article, image recognition software has to overcome
issues of viewpoint, lighting, occlusion, background,
scale, and more. The task of data augmentation prepares
these translational invariances for consideration of the
dataset such that the resulting models can perform better
despite the existing challenges. The concept of having a
larger dataset results in better deep learning models, and
performances is a generally accepted notion [50], [51].
However, collecting large datasets can be a complicated
task because of the need for manual efforts to collect and
label data, especially in the field of image processing. The
existing images from TL insulators are not an exception
to this fact, and by searching Google Images it can be
observed that there are not enough broken or defective
insulator aerial images [50].
To the best of our knowledge, there is no standard dataset
for TL insulator faults that consists of broken and
intact insulators. Therefore, because the aerial images of
insulator faults are rare and nearly impossible to collect,
to obtain adequate insulator faults images, the simulated
insulator faults samples are created on the basis of the
CPLID [52]. In these images, which are created using the
Photoshop software mentioned in [1], the normal insulator
strings are erased and replaced by their nearby pixels. The
process of using Photoshop to generate the desired dataset
is time consuming and requires considerable effort. Therefore,
only 248 faulty images are produced with this procedure
while the number of images made including intact
insulators is 3,560 [1].
Having this dataset from [1], image augmentation is the
approach that helps generate a large, sufficient dataset,
with
# += faulty insulator images and
# += intact insulator images in variant
backgrounds. The image-augmentation process is done
248 9 248 2 480
,
3 ,, ,
560 33 560 14 240
Big, Labeled Dataset
Source Model
(Feature Extraction)
Input
Source Domain
Convolutional
Layer
Fully Connected
Layers
Knowledge
(Weights of Layers)
New Dataset
(Unlabeled or Limited)
Target Model
(Feature Extraction) (Tuned Classification)
Target Task
Input
Target Domain
Convolutional
Layer
Figure 2. The schematic of a typical transfer learning model.
October 2022 IEEE SYSTEMS, MAN, & CYBERNETICS MAGAZINE 19
Fully Connected
Layers
(Classification)
Source Task
Conv_1-1
Conv_1-2
Pooling
Conv_1-1
Conv_1-2
Pooling
Conv_n-1
Conv_n-2
Pooling
Conv_n-1
Conv_n-2
Pooling
Output
Layer
Output
Layer
IEEE Systems, Man and Cybernetics Magazine - October 2022

Table of Contents for the Digital Edition of IEEE Systems, Man and Cybernetics Magazine - October 2022
